← Back
Hugging Face Blog

DaisyRose-27B: Pushing the Frontier of Post-Training Efficiency

•5 min read•
#llm#agents#deployment
DaisyRose-27B: Pushing the Frontier of Post-Training Efficiency
Level:Advanced
For:AI Engineers
✦TL;DR

DaisyRose-27B, built on Qwen3.8-27B and fine‑tuned with the Miles RL framework, doubles Terminal‑Bench 4.1 pass@1 from 5.6 % to 12.0 % after 18 days of training on 144 H200 GPUs using only 1,200 curated long‑horizon tasks. The model retains a 256K‑token context window, image understanding, and the original architecture, and is best served with SGLang + DSpark speculative decoding. Engineers can launch the model via the SGLang cookbook and query it with the OpenAI Python client, receiving both reasoning and final answer fields.

⚡ Key Takeaways

  • Terminal‑Bench 4.1 pass@1 improves from 5.6 % (Qwen3.8‑27B) to 12.0 % (DaisyRose‑27B).
  • Training used Miles, an open‑source RL framework, in fully asynchronous mode with 128 H200 GPUs for inference and 16 H200 GPUs for the trainer.
  • 57 % lower training cost compared to baseline (partial figure shown) and sessions up to 1 M tokens folded via harness summarization.
  • Recommended serving stack: SGLang with DSpark speculative decoding; launch configs provided in the SGLang cookbook.
  • API response includes `reasoning_content` (model’s chain‑of‑thought) and `content` (final answer); the model “thinks before answering” by default.
💡 Why It Matters

Doubling pass@1 on a hard coding benchmark while cutting training cost demonstrates that large‑scale RL on a tiny, high‑quality task set can yield substantial capability gains, enabling production teams to upgrade existing 27B models without prohibitive compute budgets.

✅ Practical Steps

  1. Follow the SGLang cookbook to launch DaisyRose‑27B on your GPU cluster (specify the checkpoint format and port 30000).
  2. Send a request with the OpenAI Python client to `http://localhost:30000/v1/chat/completions`, using model name `DaisyRose-27B`; parse `reasoning_content` and `content` from the response.
  3. Use Miles in fully asynchronous mode (128 inference H200 GPUs + 16 trainer H200 GPUs) for any further RL fine‑tuning on custom long‑horizon tasks.

Want the full story? Read the original article.

Read on Hugging Face Blog ↗

More like this

Introducing workload recovery and expanded infrastructure checks on Amazon SageMaker HyperPod

AWS ML Blog•#deployment

Introducing Unity Gateway Insights

Databricks Blog•#deployment

Introducing Gemini 4 Neon

Google DeepMind•#llm

Mathematical proof, powered by AI. And no hallucinations!

Amazon Science•#agents

EXPLORE AI NEWS

Daily hand-picked stories on LLMs, RAG, agents and production AI — curated for engineers who ship.

BROWSE NEWS

GET THE WEEKLY DIGEST

Join engineers getting the Monday signal-over-noise AI breakdown. No spam, unsubscribe anytime.

LEARN AI ENGINEERING

Curated courses, research papers, repos and tutorials built for engineers leveling up in AI.

START LEARNING